A double-line self-updating method and system for asphalt pavement performance estimation

Abstract

The application relates to the technical field of road maintenance, and provides a double-line self-updating method and system for asphalt pavement performance estimation, which comprises the following steps: constructing a general prediction model based on historical data and a local regression model based on measured data; through a short-term adaptive prediction path, model fusion strategy is adopted to fuse the prediction results of the two models, the weight is dynamically adjusted according to the data accumulation condition, and a final performance prediction value is output; through a long-term gradual updating path executed in parallel, new knowledge samples are detected based on a clustering model, when the samples reach a threshold value, incremental learning of the general prediction model is triggered, and self-updating of the model is realized. The application overcomes the limitations of existing static models, can dynamically adapt to different road section characteristics and data changes, and improves the accuracy, self-adaptability and long-term reliability of asphalt pavement performance estimation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of road maintenance, in particular to a double-line self-updating method and system for asphalt pavement performance estimation. BACKGROUND

[0002] With the increasing demand for road maintenance in China, data-driven asphalt pavement performance estimation has become a research focus. Currently, performance estimation models mainly rely on static machine learning models or regression models based on historical data. However, these static models are difficult to adapt to the dynamic degradation characteristics of different road sections under the influence of traffic, environment, and other factors, resulting in a significant decrease in accuracy in specific application scenarios, especially in long-term prediction. At the same time, existing models lack effective online learning mechanisms, making it difficult to fully utilize newly measured data for local correction of predictions and ensuring the robustness of predictions in the early stages of data accumulation. SUMMARY

[0003] To help solve the above technical problems, the present application provides a double-line self-updating method and system for asphalt pavement performance estimation.

[0004] In a first aspect, the present application provides a double-line self-updating method for asphalt pavement performance estimation, which adopts the following technical solution:

[0005] A double-line self-updating method for asphalt pavement performance estimation, comprising the following steps:

[0006] Model construction step: constructing a general prediction model trained based on historical road data, with historical data as input and prediction data as output, and constructing a local regression model fitted based on measured data of a target road section;

[0007] Short-term adaptive prediction path step: using a model fusion method to weight and fuse the prediction results of the general prediction model and the local regression model to obtain the final pavement performance prediction value;

[0008] Long-term gradual updating path step, executed in parallel with the short-term adaptive prediction path step, comprising:

[0009] Establishing a clustering model based on historical data distribution, calculating the deviation of a new sample corresponding to newly collected measured data of a target road section from the historical data distribution, and if the deviation exceeds a dynamic distance threshold, determining that the sample is a new knowledge sample and storing it in a sample buffer;

[0010] When the number of new knowledge samples in the sample buffer reaches a preset capacity threshold, triggering an incremental learning process of the general prediction model, updating the model using new knowledge samples in the sample buffer, and emptying the buffer after updating is completed.

[0011] Preferably, the short-term adaptive prediction path step includes:

[0012] For the target road segment, the measured pavement performance data obtained from the target road segment are collected as the fitting dataset for the local regression model. The model fusion method is adopted to dynamically calculate and allocate the fusion weights based on the prediction errors of the two models for the recent data of the target road segment.

[0013] The weighted fusion result is used as the final pavement performance prediction value for the target road segment.

[0014] Preferably, the short-term adaptive prediction path step includes:

[0015] Calculate the absolute prediction errors of the general prediction model and the local regression model for the target road segment at the most recent m time points;

[0016] Based on the absolute error, the fusion weights of the general prediction model and the local regression model are dynamically calculated and allocated using a model fusion method, wherein the weights are inversely proportional to the prediction error.

[0017] Preferably, the long-term incremental update path steps include:

[0018] Historical datasets Perform clustering, Denotes the centroid of the i-th core cluster. and radius n core clusters are obtained ;

[0019] Calculate the centroid of each core cluster and radius as follows:

[0020] ,

[0021] ,

[0022] For clusters Historical samples, Cluster The number of samples, Denotes the Euclidean norm;

[0023] Calculate new samples The smallest Euclidean distance to the centroid of all clusters Set dynamic distance threshold , The preset proportional coefficient, if > Then the new sample determining as a new knowledge sample and storing in a sample buffer.

[0024] Preferably, the long-term progressive updating path step comprises:

[0025] When the number of new knowledge samples in the sample buffer reaches a preset capacity threshold, triggering the incremental learning process of the general prediction model, updating the model using the new knowledge samples in the sample buffer, and emptying the buffer after updating. The updated general prediction model will be used for subsequent prediction and fusion in the short-term adaptive prediction path step.

[0026] In a second aspect, the application provides a double-line self-updating system for asphalt pavement performance estimation, which adopts the following technical solution:

[0027] A double-line self-updating system for asphalt pavement performance estimation, which is used to implement the double-line self-updating method for asphalt pavement performance estimation as described in any one of the first aspect, wherein it comprises:

[0028] A model construction and storage module for executing the model construction step and storing the model;

[0029] A short-term prediction module for executing the short-term adaptive prediction path step;

[0030] A long-term updating module for executing the long-term progressive updating path step;

[0031] The long-term updating module and the short-term prediction module are communicatively connected to obtain newly collected measured data, and the updated general prediction model of the long-term updating module is fed back to the model construction and storage module for calling by the short-term prediction module.

[0032] In summary, the method of the application overcomes the limitations of existing static models in asphalt pavement performance estimation through a double-line self-updating mechanism combining short-term adaptive correction and long-term progressive updating, which can dynamically adapt to the decay characteristics and data accumulation process of different road sections, improving the prediction accuracy and adaptability of the model in long-term application. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The flowchart of the double-line self-updating method for asphalt pavement performance estimation of the application;

[0034] Figure 2 The curve diagram of the prediction results of Road 1 according to the general prediction model;

[0035] Figure 3 The curve diagram of the prediction results of Road 2 according to the general prediction model;

[0036] Figure 4 Fig. 3 is a schematic diagram of a prediction result curve of the road 3 according to the prediction result of the general prediction model;

[0037] Figure 5 Fig. 4 is a schematic diagram of a result curve of the road 1, the road 2 and the road 3 based on the fitting result of the two-parameter model;

[0038] Figure 6 Fig. 5 is a schematic diagram of a three-dimensional space distribution of a historical sample clustering result and a new sample;

[0039] Figure 7 Fig. 6 is a schematic diagram of a prediction result curve of the road 2 according to the general prediction model after the general prediction model is updated. DETAILED DESCRIPTION

[0040] The application will be further described below with reference to the drawings, and the structure and principle of the application are very clear to those skilled in the art. It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application.

[0041] The two-line self-updating method for asphalt pavement performance prediction of the application comprises the following steps:

[0042] A model construction step: a general prediction model trained based on historical road data is constructed, the input is historical data, and the output is prediction data, and a local regression model fitted based on measured data of a target road section is constructed. In the embodiment of the application, the historical data can include at least one or more of current pavement performance indicators, road structure characteristics, traffic load characteristics and environmental characteristics, and the prediction data can include future pavement performance indicators, the pavement performance indicators including PCI (pavement condition index), PQI (pavement technical condition index) and RQI (road riding quality index), etc., and one of them can be selected in actual operation.

[0043] A short-term adaptive prediction path step: a model fusion strategy is adopted to perform weighted fusion on the prediction results of the general prediction model and the local regression model to obtain a final pavement performance prediction value. In the embodiment of the application, a Bayesian dynamic weighted fusion strategy can be adopted, and in other optional embodiments, a person skilled in the art can select a suitable model fusion strategy based on actual needs.

[0044] A long-term progressive updating path step, which is executed in parallel with the short-term adaptive prediction path step, comprises:

[0045] A clustering model is established based on the historical data distribution, the deviation degree of a new sample corresponding to newly collected measured data of the target road section from the historical data distribution is calculated, and if the deviation degree exceeds a dynamic distance threshold, the sample is determined as a new knowledge sample and stored in a sample buffer;

[0046] When the number of new knowledge samples in the sample buffer reaches a preset capacity threshold, an incremental learning process of the general prediction model is triggered, the model is updated using the new knowledge samples in the sample buffer, and the buffer is emptied after the update is completed.

[0047] The model construction step includes that the local regression model is a double-parameter model, which is defined by the following formula:

[0048] ,

[0049] wherein, is a current road damage index, is an initial road damage index, is a service life, and are regression parameters of the model.

[0050] It should be noted that the local regression model in the present application can also not use the above-mentioned double-parameter model but use other regression models, and the above-mentioned double-parameter model is only a specific embodiment of the present application, which aims to provide a feasible embodiment for those skilled in the art.

[0051] The short-term adaptive prediction path step includes:

[0052] For the target road section, the measured road performance data detected from the target road section is collected as a fitting data set of the local regression model, and a model fusion method is used to dynamically calculate and distribute the fusion weight according to the prediction error of the two models for the recent data of the target road section;

[0053] The weighted fusion result is taken as the final road performance prediction value of the target road section.

[0054] In order to further describe the above-mentioned "dynamically calculating and distributing the fusion weight according to the prediction error of the two models for the recent data of the target road section", the short-term adaptive prediction path step includes:

[0055] The prediction absolute errors of the general prediction model and the local regression model for the last m time points of the target road section are calculated;

[0056] Based on the absolute error, the fusion weight of the general prediction model and the local regression model is dynamically calculated and distributed by the model fusion method, wherein the weight is inversely proportional to the prediction error.

[0057] The long-term progressive update path step includes:

[0058] The historical data set is clustered, the centroid of the i-th core cluster is represented by and the radius is represented by , get n core clusters ;

[0059] Calculate the centroid of each core cluster and radius As follows:

[0060] ,

[0061] ,

[0062] For the historical samples of the cluster , Indicates the number of samples of the cluster , Indicates the Euclidean norm;

[0063] Calculate the minimum Euclidean distance of the new sample to all cluster centroids , set the dynamic distance threshold , is a preset proportion coefficient, if > The new sample is determined as a new knowledge sample and stored in a sample buffer.

[0064] In the embodiments of the present application, the value of the proportion coefficient is 1.2, and in other optional embodiments, a person skilled in the art can select a suitable proportion coefficient value based on actual needs.

[0065] The long-term progressive updating path step includes:

[0066] When the number of new knowledge samples in the sample buffer reaches the preset capacity threshold, the incremental learning process of the general prediction model is triggered, and the model is updated using the new knowledge samples in the sample buffer. After updating, the buffer is emptied, and the updated general prediction model will be used for subsequent prediction and fusion in the short-term adaptive prediction path step.

[0067] In order to further describe the above-mentioned "incremental learning process", in the long-term progressive updating path step of the embodiments of the present application, the incremental learning process adopts a second-order approximation-based parameter updating algorithm. At the same time, the general prediction model is a deep learning model based on multi-layer perception, and the MC Dropout method is used to provide uncertainty estimation for prediction. In other optional embodiments, a person skilled in the art can select a suitable incremental learning algorithm and general prediction model based on actual needs.

[0068] Specifically, the scheme of the present application proposes a double-line self-updating method for asphalt pavement performance estimation, which includes:

[0069] The first path of this mechanism employs a dynamic weighted fusion strategy, integrating and optimizing the prior knowledge of Model 1 (including but not limited to various deep learning and machine learning models, such as neural networks and multilayer perceptrons) with Model 2 (including but not limited to various regression models, such as two-parameter models and three-parameter models) based on measured data. Model 1 learns the common evolution patterns of performance indicators (including but not limited to PCI, IRI, PQI, etc.) for specific road types from historical data, while Model 2 uses measured data of road segments to fit local features. Due to the limited measured data in the initial stage, the prediction uncertainty of regression model 2 is relatively high. Therefore, the system assigns higher weights to model 1 to ensure the stability of the prediction. As measured data accumulates, the ability of regression model 2 to represent specific conditions of road segments increases. The system adjusts the weights through intelligent analysis algorithms (including but not limited to Bayesian optimization dynamic algorithms) to make the prediction results closer to the actual results. This process not only avoids the overfitting problem caused by insufficient early data but also effectively suppresses the performance evolution of historical experience models in new environments.

[0070] The second approach achieves long-term adaptive updates for Model 1 through new knowledge detection and progressive retraining. When the actual performance evolution trend of road segments differs significantly from historical experience, short-term correction strategies may not completely eliminate prediction bias. Therefore, the system introduces a new knowledge sample detection module, which identifies samples deviating from the historical distribution through an anomaly detection mechanism (based on clustering algorithms and solving the Euclidean norm of samples). When a sufficient number of anomaly samples are detected, the incremental learning mechanism of Model 1 is triggered, incorporating new data into the training set and employing intelligent algorithms (including but not limited to second-order approximate parameter update algorithms) for incremental learning to avoid catastrophic forgetting. This progressive update mechanism reduces the computational cost of frequent retraining while ensuring that the model absorbs new patterns while retaining historical knowledge, thus adapting to the dynamic evolution of long-term performance degradation patterns.

[0071] The above anomaly detection mechanism is as follows:

[0072] Historical datasets Perform clustering, Denotes the centroid of the i-th core cluster. and radius n core clusters are obtained Based on this, a new knowledge detection function with a dynamic threshold is designed. If a new sample exceeds the trigger threshold τ, it will be identified as a new knowledge sample and stored in the new knowledge buffer. When the number of samples in the new knowledge buffer exceeds the quantity threshold, an incremental learning process based on second-order approximation will be activated.

[0073] The new knowledge detection function is as follows:

[0074] Historical datasets perform clustering, denote the centroid of the i-th core cluster and radius , obtaining n core clusters ;

[0075] calculate the centroid and radius of each core cluster as follows:

[0076] ,

[0077] ,

[0078] for the historical samples of cluster , denote the number of samples of cluster , denote the Euclidean norm;

[0079] calculate the minimum Euclidean distance of new sample to all cluster centroids set the dynamic distance threshold , is a preset proportion coefficient, if > then the new sample is determined as a new knowledge sample and stored in a sample buffer.

[0080] The above-mentioned "triggering the incremental learning mechanism of model 1" has the following rules:

[0081] When the number of samples in the buffer reaches the capacity threshold of the sample buffer, the system will automatically trigger the update mechanism of model 1. In this process, the new knowledge samples in the buffer will be converted into historical experience data, and then the sample buffer will be emptied for subsequent use.

[0082] In actual operation, the proposed double-line self-updating mechanism is used to predict the road performance index PCI of three roads in Shanghai:

[0083] The original data of the three roads are as follows in Table 1:

[0084]

[0085] (1) Construct model 1 based on historical experience using multilayer perceptron (MLP) + MC Dropout algorithm:

[0086] ​The input of the adopted MLP model is 55-dimensional, including PCI and five categories of influence factor indexes of structure, road, traffic, environment and maintenance, in addition to feature missing mask, and the output is the PCI value of future n years. The model can effectively capture the evolution law of PCI under the action of natural decay and maintenance intervention, and the MC Dropout method further enhances the robustness of the model, so that it can maintain good prediction performance under incomplete data conditions and give prediction mean and confidence interval.

[0087] Since the case data only contains 5 features, there are a large number of data missing, so the prediction result given by model 1 based on historical experience has large error. The prediction results of the three roads according to model 1 are shown in Figure 2 、 Figure 3 、 Figure 4

[0088] (2) Model 2 based on measured data is constructed by using a two-parameter model:

[0089] The two-parameter model is defined by the following formula:

[0090] ,

[0091] wherein, is the current pavement damage index, is the initial pavement damage index, is the service life, and are regression parameters of the model.

[0092] The fitting result of the two-parameter model is shown in Figure 5

[0093] (3) Short-term adaptive correction of model 1 and model 2:

[0094] Bayesian dynamic weighting is used to integrate the MC Dropout corrected MLP model and the two-parameter regression model. This method can dynamically adjust the model weight according to the calculation error of the two models, and the prediction results of the model after integrated learning are as shown in Table 2:

[0095]

[0096] ​​It can be seen that the prediction accuracy of the model fused by the ensemble learning method is significantly improved. This method innovatively combines the advantages of the double-parameter regression model and the MLP model: on the one hand, it inherits the ability to accurately capture the long-term decay trend of the double-parameter regression model, and on the other hand, it fully utilizes the robust prediction advantages of the MLP model under data constraints such as insufficient data accumulation and missing values. This complementary fusion strategy not only improves the prediction accuracy, but also enhances the adaptability of the model in practical engineering applications.

[0097] (4) Long-term progressive learning through new empirical data:

[0098] First, the data is analyzed by the new knowledge sample detection module. In order to facilitate visual display, principal component analysis (PCA) is used to reduce the original 55-dimensional feature space to 3 dimensions, and then the DBSCAN density clustering algorithm is used to divide the samples into 4 clusters, as shown in Figure 6 By calculating the distance of the 57 new samples generated by the three measured road segments to the nearest cluster center, it is found that the data samples of road 2 are significantly higher than the preset threshold, so they are judged as new knowledge samples and stored in the sample buffer. To verify the self-updating effect of the algorithm, this embodiment breaks through the conventional process limit and directly updates the MLP model weights using the incremental learning algorithm based on the second-order approximation when the sample buffer does not reach the capacity threshold.

[0099] After model updating, the prediction results of model 1 for road 2 are as shown in Figure 7 Compared with the original prediction results of Figure 3 , it can be obviously observed that the prediction accuracy and uncertainty quantification are significantly improved after model updating. Specifically, in terms of prediction error, the updated model reduces the average prediction error of road 2 from 9.9 to 2.68, and the maximum error from 18.0 to 6.60; in terms of confidence interval, the average width is reduced from 68.2 to 24.24, and the maximum width is significantly reduced from 100 to 36.16. This result shows that the incremental learning algorithm proposed in this study effectively improves the model's ability to adapt to new knowledge. By identifying the data of road 2 as new knowledge samples and performing targeted model updating, not only the prediction accuracy of this road segment is significantly improved, but also the uncertainty of prediction is greatly reduced.

[0100] The present application also proposes a double-line self-updating system for asphalt pavement performance estimation, which is used for the double-line self-updating method for asphalt pavement performance estimation described above, comprising:

[0101] a model construction and storage module for executing the model construction step and storing the model;

[0102] a short-term prediction module for executing the short-term adaptive prediction path step;

[0103] a long-term updating module for performing a long-term progressive updating path step;

[0104] The long-term updating module is in communication connection with the short-term prediction module to obtain newly collected measured data, and the updated general prediction model of the long-term updating module is fed back to the model construction and storage module for calling by the short-term prediction module.

Claims

1. A double-line self-updating method for asphalt pavement performance estimation, characterized in that, The method comprises the following steps: A model construction step: constructing a general prediction model trained based on historical road data, with input being historical data and output being prediction data, and constructing a local regression model fitted based on measured data of a target road section; A short-term adaptive prediction path step: using a model fusion method to perform weighted fusion on prediction results of the general prediction model and the local regression model, to obtain a final road surface performance prediction value; A long-term progressive updating path step, which is executed in parallel with the short-term adaptive prediction path step, comprising: Establishing a clustering model based on historical data distribution, calculating a deviation degree of a new sample corresponding to newly collected measured data of a target road section from the historical data distribution, and determining that the sample is a new knowledge sample and storing it in a sample buffer if the deviation degree exceeds a dynamic distance threshold; When the number of new knowledge samples in the sample buffer reaches a preset capacity threshold, triggering an incremental learning process of the general prediction model, updating the model using the new knowledge samples in the sample buffer, emptying the buffer after the updating is completed, and using the updated general prediction model for prediction and fusion in the short-term adaptive prediction path step; The short-term adaptive prediction path step comprises: For a target road section, collecting measured road surface performance data detected from the target road section as a fitting data set of the local regression model, and using a model fusion method to dynamically calculate and assign fusion weights of the two models according to prediction errors of the two models on recent data of the target road section; Taking the weighted fusion result as a final road surface performance prediction value of the target road section.

2. The double line self-refresh method for asphalt pavement performance prediction according to claim 1, characterized in that, The short-term adaptive prediction path step comprises: Calculating prediction absolute errors of the general prediction model and the local regression model on the target road section at the last m time points; Based on the absolute errors, dynamically calculating and assigning fusion weights of the general prediction model and the local regression model by a model fusion method, wherein the weights are inversely proportional to the prediction absolute errors.

3. The double line self-refresh method for asphalt pavement performance prediction according to claim 1, wherein, The long-term progressive updating path step comprises: clustering a historical data set performing clustering on the historical data set, representing a centroid of the i-th core cluster and a radius resulting in n core clusters ; Compute the centroid of each core cluster and radius As follows: ， ， is a cluster of historical samples, represents the number of samples of the cluster , and denotes the Euclidean norm; Compute new sample Minimum Euclidean distance to all cluster centroids Set dynamic distance threshold , is a preset proportion coefficient, if > The new sample is determined as a new knowledge sample and stored in a sample buffer.

4. A double-line self-refresh system for asphalt pavement performance estimation, for implementing the double-line self-refresh method for asphalt pavement performance estimation according to any one of claims 1 to 3, characterized in that, Comprising: A model construction and storage module for executing the model construction step and storing the model; A short-term prediction module for executing the short-term adaptive prediction path step; A long-term updating module for executing the long-term progressive updating path step; The long-term updating module and the short-term prediction module are communicatively connected to obtain newly collected measured data; and the updated general prediction model of the long-term updating module is fed back to the model construction and storage module for calling by the short-term prediction module.

Images